def _setup_policy(self, time_step_spec, action_spec, boltzmann_temperature, emit_log_probability): policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._q_network, emit_log_probability=emit_log_probability, observation_and_action_constraint_splitter=( self._observation_and_action_constraint_splitter)) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) # Create self._target_greedy_policy in order to compute target Q-values. target_policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._target_q_network, observation_and_action_constraint_splitter=( self._observation_and_action_constraint_splitter)) self._target_greedy_policy = greedy_policy.GreedyPolicy(target_policy) return policy, collect_policy
def _get_policies(self, time_step_spec, action_spec, cloning_network): policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._cloning_network) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) return policy, collect_policy
def _setup_policy(self, time_step_spec, action_spec, boltzmann_temperature, emit_log_probability): policy = categorical_q_policy.CategoricalQPolicy( time_step_spec, action_spec, self._q_network, self._min_q_value, self._max_q_value, observation_and_action_constraint_splitter=( self._observation_and_action_constraint_splitter)) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) target_policy = categorical_q_policy.CategoricalQPolicy( time_step_spec, action_spec, self._target_q_network, self._min_q_value, self._max_q_value, observation_and_action_constraint_splitter=( self._observation_and_action_constraint_splitter)) self._target_greedy_policy = greedy_policy.GreedyPolicy(target_policy) return policy, collect_policy
def testTensorEpsilon(self): epsilon_ph = tf.placeholder(tf.float32, shape=()) policy = epsilon_greedy_policy.EpsilonGreedyPolicy(self._policy, epsilon=epsilon_ph) self.assertEqual(policy.time_step_spec(), self._time_step_spec) self.assertEqual(policy.action_spec(), self._action_spec) policy_state = policy.get_initial_state(batch_size=2) action_step = policy.action(self._time_step, policy_state, seed=54) nest.assert_same_structure(self._action_spec, action_step.action) self.evaluate(tf.global_variables_initializer()) with self.cached_session() as sess: for epsilon in [0.0, 0.2, 0.7, 1.0]: # Collect 100 steps with the current value of epsilon. actions = [] num_steps = 1000 for _ in range(num_steps): action_ = sess.run(action_step.action, {epsilon_ph: epsilon})[0] self.assertIn(action_, [0, 1, 2]) actions.append(action_) # Verify that action distribution changes as we vary epsilon. self.checkActionDistribution(actions, epsilon, num_steps)
def testTensorEpsilon(self, epsilon): policy = epsilon_greedy_policy.EpsilonGreedyPolicy( self._policy, epsilon=epsilon) self.assertEqual(policy.time_step_spec, self._time_step_spec) self.assertEqual(policy.action_spec, self._action_spec) policy_state = policy.get_initial_state(batch_size=2) time_step = tf.nest.map_structure(tf.convert_to_tensor, self._time_step) @common.function def action_step_fn(time_step=time_step): return policy.action(time_step, policy_state, seed=54) tf.nest.assert_same_structure( self._action_spec, self.evaluate(action_step_fn(time_step)).action) if tf.executing_eagerly(): action_step = action_step_fn else: action_step = action_step_fn() actions = [] num_steps = 1000 for _ in range(num_steps): action_ = self.evaluate(action_step).action[0] self.assertIn(action_, [0, 1, 2]) actions.append(action_) # Verify that action distribution changes as we vary epsilon. self.checkActionDistribution(actions, epsilon, num_steps)
def __init__( self, time_step_spec, action_spec, reward_network, optimizer, epsilon, # Params for training. error_loss_fn=tf.compat.v1.losses.mean_squared_error, gradient_clipping=None, # Params for debugging. debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a Neural Epsilon Greedy Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. reward_network: A `tf_agents.network.Network` to be used by the agent. The network will be called with call(observation, step_type) and it is expected to provide a reward prediction for all actions. optimizer: The optimizer to use for training. epsilon: A float representing the probability of choosing a random action instead of the greedy action. error_loss_fn: A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is `tf.losses.mean_squared_error`. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0. """ super(NeuralEpsilonGreedyAgent, self).__init__(time_step_spec=time_step_spec, action_spec=action_spec, reward_network=reward_network, optimizer=optimizer, observation_and_action_constraint_splitter=None, error_loss_fn=error_loss_fn, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter, name=name) self._policy = epsilon_greedy_policy.EpsilonGreedyPolicy( self._policy, epsilon=epsilon) self._collect_policy = self._policy
def get_policy(env, q_net, epsilon_callback): q_plcy = q_policy.QPolicy(env.time_step_spec(), env.action_spec(), q_network=q_net) # greedy_plcy = greedy_policy.GreedyPolicy(q_plcy) ep_greedy_plcy = epsilon_greedy_policy.EpsilonGreedyPolicy( q_plcy, epsilon_callback) plcy = ep_greedy_plcy return plcy
def testInfoFromGreedy(self): PolicyInfo = collections.namedtuple( # pylint: disable=invalid-name 'PolicyInfo', ('log_probability', 'predicted_rewards', 'bandit_policy_type')) # Set default empty tuple for all fields. PolicyInfo.__new__.__defaults__ = ((), ) * len(PolicyInfo._fields) info_spec = PolicyInfo( bandit_policy_type=self._bandit_policy_type_spec, log_probability=tensor_spec.BoundedTensorSpec( shape=(), dtype=tf.float32, maximum=0, minimum=-float('inf'), name='log_probability')) policy_with_info_spec = fixed_policy.FixedPolicy( np.asarray(self._greedy_action, dtype=np.int32), self._time_step_spec, self._action_spec, policy_info=PolicyInfo( bandit_policy_type=self._bandit_policy_type), info_spec=info_spec) epsilon = 0.2 policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy_with_info_spec, epsilon=epsilon, info_fields_to_inherit_from_greedy=['log_probability']) self.assertEqual(policy.time_step_spec, self._time_step_spec) self.assertEqual(policy.action_spec, self._action_spec) time_step = tf.nest.map_structure(tf.convert_to_tensor, self._time_step) @common.function def action_step_fn(time_step=time_step): return policy.action(time_step, policy_state=(), seed=54) tf.nest.assert_same_structure( self._action_spec, self.evaluate(action_step_fn(time_step)).action) if tf.executing_eagerly(): action_step = action_step_fn else: action_step = action_step_fn() step = self.evaluate(action_step) tf.nest.assert_same_structure(info_spec, step.info) self.checkBanditPolicyTypeShape(step.info.bandit_policy_type, batch_size=2) self.assertAllEqual(step.info.log_probability, tf.zeros_like(step.info.log_probability))
def _get_policies(self, time_step_spec, action_spec, cloning_network): policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._cloning_network, # Unlike DQN, we support continuous action spaces - in which case # the policy just emits the network output. In that case, we # don't care if the action_spec is a scalar integer value. validate_action_spec_and_network=False, ) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) return policy, collect_policy
def _setup_as_discrete(self, time_step_spec, action_spec, loss_fn, epsilon_greedy): self._loss_fn = loss_fn or self._discrete_loss # Unlike DQN, we support continuous action spaces - in which case # the policy just emits the network output. In that case, we # don't care if the action_spec is a scalar integer value. policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._cloning_network, validate_action_spec_and_network=False, ) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) return policy, collect_policy
def _setup_policy(self, time_step_spec, action_spec, boltzmann_temperature, emit_log_probability): policy = q_policy.QPolicy(time_step_spec, action_spec, q_network=self._q_network, emit_log_probability=emit_log_probability) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) return policy, collect_policy
def _setup_policy(self, time_step_spec, action_spec, emit_log_probability): policy = qtopt_cem_policy.CEMPolicy( time_step_spec, action_spec, q_network=self._target_q_network, sampler=self._sampler, init_mean=self._init_mean_cem, init_var=self._init_var_cem, info_spec=self._info_spec, num_samples=self._num_samples_cem, num_elites=self._num_elites_cem, num_iterations=self._num_iter_cem, emit_log_probability=emit_log_probability, training=False) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) return policy, collect_policy
def testFixedEpsilon(self, epsilon): policy = epsilon_greedy_policy.EpsilonGreedyPolicy(self._policy, epsilon=epsilon) self.assertEqual(policy.time_step_spec(), self._time_step_spec) self.assertEqual(policy.action_spec(), self._action_spec) policy_state = policy.get_initial_state(batch_size=2) action_step = policy.action(self._time_step, policy_state, seed=54) nest.assert_same_structure(self._action_spec, action_step.action) self.evaluate(tf.global_variables_initializer()) # Collect 100 steps with the current value of epsilon. actions = [] num_steps = 100 for _ in range(num_steps): action_ = self.evaluate(action_step.action)[0] self.assertIn(action_, [0, 1, 2]) actions.append(action_) self.checkActionDistribution(actions, epsilon, num_steps)
def _setup_as_discrete(self, time_step_spec, action_spec, loss_fn, epsilon_greedy): self._bc_loss_fn = loss_fn or self._discrete_loss if any(isinstance(d, distribution_utils.DistributionSpecV2) for d in tf.nest.flatten([self._network_output_spec])): # If the output of the cloning network contains a distribution. base_policy = actor_policy.ActorPolicy(time_step_spec, action_spec, self._cloning_network) else: # If the output of the cloning network is logits. base_policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._cloning_network, validate_action_spec_and_network=False) policy = greedy_policy.GreedyPolicy(base_policy) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( base_policy, epsilon=epsilon_greedy) return policy, collect_policy
def testInfoSpec(self): PolicyInfo = collections.namedtuple( # pylint: disable=invalid-name 'PolicyInfo', ('log_probability', 'predicted_rewards')) # Set default empty tuple for all fields. PolicyInfo.__new__.__defaults__ = ((), ) * len(PolicyInfo._fields) info_spec = PolicyInfo() policy_with_info_spec = fixed_policy.FixedPolicy( np.asarray([self._greedy_action], dtype=np.int32), self._time_step_spec, self._action_spec, policy_info=PolicyInfo(), info_spec=info_spec) epsilon = 0.2 policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy_with_info_spec, epsilon=epsilon) self.assertEqual(policy.time_step_spec, self._time_step_spec) self.assertEqual(policy.action_spec, self._action_spec) time_step = tf.nest.map_structure(tf.convert_to_tensor, self._time_step) @common.function def action_step_fn(time_step=time_step): return policy.action(time_step, policy_state=(), seed=54) tf.nest.assert_same_structure( self._action_spec, self.evaluate(action_step_fn(time_step)).action) if tf.executing_eagerly(): action_step = action_step_fn else: action_step = action_step_fn() step = self.evaluate(action_step) tf.nest.assert_same_structure(info_spec, step.info)
def __init__( self, time_step_spec, action_spec, categorical_q_network, optimizer, min_q_value=-10.0, max_q_value=10.0, epsilon_greedy=0.1, n_step_update=1, boltzmann_temperature=None, # Params for target network updates target_update_tau=1.0, target_update_period=1, # Params for training. td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, gradient_clipping=None, # Params for debugging debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a Categorical DQN Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A `BoundedTensorSpec` representing the actions. categorical_q_network: A categorical_q_network.CategoricalQNetwork that returns the q_distribution for each action. optimizer: The optimizer to use for training. min_q_value: A float specifying the minimum Q-value, used for setting up the support. max_q_value: A float specifying the maximum Q-value, used for setting up the support. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). n_step_update: The number of steps to consider when computing TD error and TD loss. Defaults to single-step updates. Note that this requires the user to call train on Trajectory objects with a time dimension of `n_step_update + 1`. However, note that we do not yet support `n_step_update > 1` in the case of RNNs (i.e., non-empty `q_network.state_spec`). boltzmann_temperature: Temperature value to use for Boltzmann sampling of the actions during data collection. The closer to 0.0, the higher the probability of choosing the best action. target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of element_wise_huber_loss is used. This function takes as input the target and the estimated Q values and returns the loss for each element of the batch. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: TypeError: If the action spec contains more than one action. """ num_atoms = getattr(categorical_q_network, 'num_atoms', None) if num_atoms is None: raise TypeError( 'Expected categorical_q_network to have property ' '`num_atoms`, but it doesn\'t (note: you likely want to ' 'use a CategoricalQNetwork). Network is: %s' % (categorical_q_network, )) self._num_atoms = num_atoms self._min_q_value = min_q_value self._max_q_value = max_q_value self._support = tf.linspace(min_q_value, max_q_value, num_atoms) super(CategoricalDqnAgent, self).__init__(time_step_spec, action_spec, categorical_q_network, optimizer, epsilon_greedy=epsilon_greedy, n_step_update=n_step_update, boltzmann_temperature=boltzmann_temperature, target_update_tau=target_update_tau, target_update_period=target_update_period, td_errors_loss_fn=td_errors_loss_fn, gamma=gamma, reward_scale_factor=reward_scale_factor, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter, name=name) policy = categorical_q_policy.CategoricalQPolicy( min_q_value, max_q_value, self._q_network, self._action_spec) if boltzmann_temperature is not None: self._collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: self._collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) self._policy = greedy_policy.GreedyPolicy(policy)
def train(): summary_interval = 1000 summaries_flush_secs = 10 num_eval_episodes = 5 root_dir = '/tmp/tensorflow/logs/tfenv01' train_dir = os.path.join(root_dir, 'train') eval_dir = os.path.join(root_dir, 'eval') train_summary_writer = tf.compat.v2.summary.create_file_writer( train_dir, flush_millis=summaries_flush_secs*1000) train_summary_writer.set_as_default() eval_summary_writer = tf.compat.v2.summary.create_file_writer( eval_dir, flush_millis=summaries_flush_secs*1000) # maybe py_metrics? eval_metrics = [ tf_metrics.AverageReturnMetric(buffer_size=num_eval_episodes), tf_metrics.AverageEpisodeLengthMetric(buffer_size=num_eval_episodes), ] environment = TradeEnvironment() # utils.validate_py_environment(environment, episodes=5) # Environments global_step = tf.compat.v1.train.get_or_create_global_step() with tf.compat.v2.summary.record_if( lambda: tf.math.equal(global_step % summary_interval, 0)): train_env = tf_py_environment.TFPyEnvironment(environment) eval_env = tf_py_environment.TFPyEnvironment(environment) num_iterations = 50 fc_layer_params = (512, ) # ~ (17 + 1001) / 2 input_fc_layer_params = (50, ) output_fc_layer_params = (20, ) lstm_size = (30, ) initial_collect_steps = 20 collect_steps_per_iteration = 1 collect_episodes_per_iteration = 1 # the same as above batch_size = 64 replay_buffer_capacity = 10000 train_sequence_length = 10 gamma = 0.99 # check if 1.0 works as well target_update_tau = 0.05 target_update_period = 5 epsilon_greedy = 0.1 gradient_clipping = None reward_scale_factor = 1.0 learning_rate = 1e-2 log_interval = 30 eval_interval = 15 # train_env.observation_spec(), q_net = q_rnn_network.QRnnNetwork( train_env.time_step_spec().observation, train_env.action_spec(), input_fc_layer_params=input_fc_layer_params, lstm_size=lstm_size, output_fc_layer_params=output_fc_layer_params, ) optimizer = tf.compat.v1.train.AdamOptimizer( learning_rate=learning_rate) tf_agent = dqn_agent.DqnAgent( train_env.time_step_spec(), train_env.action_spec(), q_network=q_net, optimizer=optimizer, epsilon_greedy=epsilon_greedy, target_update_tau=target_update_tau, target_update_period=target_update_period, td_errors_loss_fn=dqn_agent.element_wise_squared_loss, gamma=gamma, reward_scale_factor=reward_scale_factor, gradient_clipping=gradient_clipping, debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=global_step, ) replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer( tf_agent.collect_data_spec, batch_size=train_env.batch_size, max_length=replay_buffer_capacity, ) train_metrics = [ tf_metrics.NumberOfEpisodes(), tf_metrics.EnvironmentSteps(), tf_metrics.AverageReturnMetric(), tf_metrics.AverageEpisodeLengthMetric(), ] # Policy which does not allow some actions in certain states q_policy = FilteredQPolicy( tf_agent._time_step_spec, tf_agent._action_spec, q_network=tf_agent._q_network, ) # Valid policy to pre-fill replay buffer initial_collect_policy = DummyTradePolicy( train_env.time_step_spec(), train_env.action_spec(), ) print('Initial collecting...') initial_collect_op = dynamic_episode_driver.DynamicEpisodeDriver( train_env, initial_collect_policy, observers=[replay_buffer.add_batch] + train_metrics, num_episodes=initial_collect_steps, ).run() # Main agent's policy; greedy one policy = greedy_policy.GreedyPolicy(q_policy) # Policy used for evaluation, the same as above eval_policy = greedy_policy.GreedyPolicy(q_policy) tf_agent._policy = policy collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( q_policy, epsilon=tf_agent._epsilon_greedy) # Patch random policy for epsilon greedy collect policy filtered_random_tf_policy = FilteredRandomTFPolicy( time_step_spec=policy.time_step_spec, action_spec=policy.action_spec, ) collect_policy._random_policy = filtered_random_tf_policy tf_agent._collect_policy = collect_policy collect_op = dynamic_episode_driver.DynamicEpisodeDriver( train_env, collect_policy, observers=[replay_buffer.add_batch] + train_metrics, num_episodes=collect_episodes_per_iteration, ).run() dataset = replay_buffer.as_dataset( num_parallel_calls=3, sample_batch_size=batch_size, num_steps=train_sequence_length+1, ).prefetch(3) iterator = iter(dataset) experience, _ = next(iterator) loss_info = common.function(tf_agent.train)(experience=experience) # Checkpoints train_checkpointer = common.Checkpointer( ckpt_dir=train_dir, agent=tf_agent, global_step=global_step, metrics=metric_utils.MetricsGroup(train_metrics, 'train_metrics'), ) policy_checkpointer = common.Checkpointer( ckpt_dir=os.path.join(train_dir, 'policy'), policy=tf_agent.policy, global_step=global_step, ) rb_checkpointer = common.Checkpointer( ckpt_dir=os.path.join(train_dir, 'replay_buffer'), max_to_keep=1, replay_buffer=replay_buffer, ) summary_ops = [] for train_metric in train_metrics: summary_ops.append(train_metric.tf_summaries( train_step=global_step, step_metrics=train_metrics[:2], )) with eval_summary_writer.as_default(), \ tf.compat.v2.summary.record_if(True): for eval_metric in eval_metrics: eval_metric.tf_summaries(train_step=global_step) init_agent_op = tf_agent.initialize() with tf.compat.v1.Session() as sess: # sess.run(train_summary_writer.init()) # sess.run(eval_summary_writer.init()) # Initialize the graph # tfe.Saver().restore() # train_checkpointer.initialize_or_restore() # rb_checkpointer.initialize_or_restore() # sess.run(iterator.initializer) common.initialize_uninitialized_variables(sess) sess.run(init_agent_op) print('Collecting initial experience...') sess.run(initial_collect_op) global_step_val = sess.run(global_step) metric_utils.compute_summaries( eval_metrics, eval_env, eval_policy, num_episodes=num_eval_episodes, global_step=global_step_val, callback=eval_metrics_callback, log=True, ) collect_call = sess.make_callable(collect_op) train_step_call = sess.make_callable([loss_info, summary_ops]) global_step_call = sess.make_callable(global_step) timed_at_step = global_step_call() time_acc = 0 steps_per_second_ph = tf.compat.v1.placeholder( tf.float32, shape=(), name='steps_per_sec_ph') steps_per_second_summary = tf.compat.v2.summary.scalar( name='global_steps_per_sec', data=steps_per_second_ph, step=global_step, ) # Train for i in range(num_iterations): start_time = time.time() collect_call() for _ in range(train_steps_per_iteration): loss_info_value, _ = train_step_call() time_acc += time.time() - start_time global_step_val = global_step_call() if global_step_val % log_inerval == 0: print('step=%d, loss=%f', global_step_val, loss_info_value.loss) steps_per_sec = (global_step_val-timed_at_step) / time_acc print('%.3f steps/sec', steps_per_sec) sess.run( steps_per_second_summary, feed_dict={steps_per_second_ph: steps_per_sec}, ) timed_at_step = global_step_val time_acc = 0 # Save checkpoints if global_step_val % train_checkpoint_interval == 0: train_checkpointer.save(global_step=global_step_val) if global_step_val % policy_checkpoint_interval == 0: policy_checkpointer.save(global_step=global_step_val) if global_step_val % rb_checkpoint_interval == 0: rb_checkpointer.save(global_step=global_step_val) # Evaluate if global_step_val % eval_interval == 0: metric_utils.compute_summaries( eval_metrics, eval_env, eval_policy, num_episodes=num_eval_episodes, global_step=global_step_val, log=True, callback=eval_metrics_callback, ) print('Done!')
def __init__( self, root_dir, env_name, num_iterations=200, max_episode_frames=108000, # ALE frames terminal_on_life_loss=False, conv_layer_params=((32, (8, 8), 4), (64, (4, 4), 2), (64, (3, 3), 1)), fc_layer_params=(512, ), # Params for collect initial_collect_steps=80000, # ALE frames epsilon_greedy=0.01, epsilon_decay_period=1000000, # ALE frames replay_buffer_capacity=1000000, # Params for train train_steps_per_iteration=1000000, # ALE frames update_period=16, # ALE frames target_update_tau=1.0, target_update_period=32000, # ALE frames batch_size=32, learning_rate=2.5e-4, n_step_update=2, gamma=0.99, reward_scale_factor=1.0, gradient_clipping=None, # Params for eval do_eval=True, eval_steps_per_iteration=500000, # ALE frames eval_epsilon_greedy=0.001, # Params for checkpoints, summaries, and logging log_interval=1000, summary_interval=1000, summaries_flush_secs=10, debug_summaries=True, summarize_grads_and_vars=True, eval_metrics_callback=None): """A simple Atari train and eval for DQN. Args: root_dir: Directory to write log files to. env_name: Fully-qualified name of the Atari environment (i.e. Pong-v0). num_iterations: Number of train/eval iterations to run. max_episode_frames: Maximum length of a single episode, in ALE frames. terminal_on_life_loss: Whether to simulate an episode termination when a life is lost. conv_layer_params: Params for convolutional layers of QNetwork. fc_layer_params: Params for fully connected layers of QNetwork. initial_collect_steps: Number of frames to ALE frames to process before beginning to train. Since this is in ALE frames, there will be initial_collect_steps/4 items in the replay buffer when training starts. epsilon_greedy: Final epsilon value to decay to for training. epsilon_decay_period: Period over which to decay epsilon, from 1.0 to epsilon_greedy (defined above). replay_buffer_capacity: Maximum number of items to store in the replay buffer. train_steps_per_iteration: Number of ALE frames to run through for each iteration of training. update_period: Run a train operation every update_period ALE frames. target_update_tau: Coeffecient for soft target network updates (1.0 == hard updates). target_update_period: Period, in ALE frames, to copy the live network to the target network. batch_size: Number of frames to include in each training batch. learning_rate: RMS optimizer learning rate. n_step_update: The number of steps to consider when computing TD error and TD loss. Applies standard single-step updates when set to 1. gamma: Discount for future rewards. reward_scale_factor: Scaling factor for rewards. gradient_clipping: Norm length to clip gradients. do_eval: If True, run an eval every iteration. If False, skip eval. eval_steps_per_iteration: Number of ALE frames to run through for each iteration of evaluation. eval_epsilon_greedy: Epsilon value to use for the evaluation policy (0 == totally greedy policy). log_interval: Log stats to the terminal every log_interval training steps. summary_interval: Write TF summaries every summary_interval training steps. summaries_flush_secs: Flush summaries to disk every summaries_flush_secs seconds. debug_summaries: If True, write additional summaries for debugging (see dqn_agent for which summaries are written). summarize_grads_and_vars: Include gradients in summaries. eval_metrics_callback: A callback function that takes (metric_dict, global_step) as parameters. Called after every eval with the results of the evaluation. """ self._update_period = update_period / ATARI_FRAME_SKIP self._train_steps_per_iteration = (train_steps_per_iteration / ATARI_FRAME_SKIP) self._do_eval = do_eval self._eval_steps_per_iteration = eval_steps_per_iteration / ATARI_FRAME_SKIP self._eval_epsilon_greedy = eval_epsilon_greedy self._initial_collect_steps = initial_collect_steps / ATARI_FRAME_SKIP self._summary_interval = summary_interval self._num_iterations = num_iterations self._log_interval = log_interval self._eval_metrics_callback = eval_metrics_callback with gin.unlock_config(): gin.bind_parameter(('tf_agents.environments.atari_preprocessing.' 'AtariPreprocessing.terminal_on_life_loss'), terminal_on_life_loss) root_dir = os.path.expanduser(root_dir) train_dir = os.path.join(root_dir, 'train') eval_dir = os.path.join(root_dir, 'eval') train_summary_writer = tf.compat.v2.summary.create_file_writer( train_dir, flush_millis=summaries_flush_secs * 1000) train_summary_writer.set_as_default() self._train_summary_writer = train_summary_writer self._eval_summary_writer = None if self._do_eval: self._eval_summary_writer = tf.compat.v2.summary.create_file_writer( eval_dir, flush_millis=summaries_flush_secs * 1000) self._eval_metrics = [ py_metrics.AverageReturnMetric(name='PhaseAverageReturn', buffer_size=np.inf), py_metrics.AverageEpisodeLengthMetric( name='PhaseAverageEpisodeLength', buffer_size=np.inf), ] self._global_step = tf.compat.v1.train.get_or_create_global_step() with tf.compat.v2.summary.record_if(lambda: tf.math.equal( self._global_step % self._summary_interval, 0)): self._env = suite_atari.load( env_name, max_episode_steps=max_episode_frames / ATARI_FRAME_SKIP, gym_env_wrappers=suite_atari. DEFAULT_ATARI_GYM_WRAPPERS_WITH_STACKING) self._env = batched_py_environment.BatchedPyEnvironment( [self._env]) observation_spec = tensor_spec.from_spec( self._env.observation_spec()) time_step_spec = ts.time_step_spec(observation_spec) action_spec = tensor_spec.from_spec(self._env.action_spec()) with tf.device('/cpu:0'): epsilon = tf.compat.v1.train.polynomial_decay( 1.0, self._global_step, epsilon_decay_period / ATARI_FRAME_SKIP / self._update_period, end_learning_rate=epsilon_greedy) with tf.device('/gpu:0'): optimizer = tf.compat.v1.train.RMSPropOptimizer( learning_rate=learning_rate, decay=0.95, momentum=0.0, epsilon=0.00001, centered=True) categorical_q_net = AtariCategoricalQNetwork( observation_spec, action_spec, conv_layer_params=conv_layer_params, fc_layer_params=fc_layer_params) agent = categorical_dqn_agent.CategoricalDqnAgent( time_step_spec, action_spec, categorical_q_network=categorical_q_net, optimizer=optimizer, epsilon_greedy=epsilon, n_step_update=n_step_update, target_update_tau=target_update_tau, target_update_period=(target_update_period / ATARI_FRAME_SKIP / self._update_period), gamma=gamma, reward_scale_factor=reward_scale_factor, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=self._global_step) self._collect_policy = py_tf_policy.PyTFPolicy( agent.collect_policy) if self._do_eval: self._eval_policy = py_tf_policy.PyTFPolicy( epsilon_greedy_policy.EpsilonGreedyPolicy( policy=agent.policy, epsilon=self._eval_epsilon_greedy)) py_observation_spec = self._env.observation_spec() py_time_step_spec = ts.time_step_spec(py_observation_spec) py_action_spec = policy_step.PolicyStep( self._env.action_spec()) data_spec = trajectory.from_transition(py_time_step_spec, py_action_spec, py_time_step_spec) self._replay_buffer = py_hashed_replay_buffer.PyHashedReplayBuffer( data_spec=data_spec, capacity=replay_buffer_capacity) with tf.device('/cpu:0'): ds = self._replay_buffer.as_dataset( sample_batch_size=batch_size, num_steps=n_step_update + 1) ds = ds.prefetch(4) ds = ds.apply( tf.data.experimental.prefetch_to_device('/gpu:0')) with tf.device('/gpu:0'): self._ds_itr = tf.compat.v1.data.make_one_shot_iterator(ds) experience = self._ds_itr.get_next() self._train_op = agent.train(experience) self._env_steps_metric = py_metrics.EnvironmentSteps() self._step_metrics = [ py_metrics.NumberOfEpisodes(), self._env_steps_metric, ] self._train_metrics = self._step_metrics + [ py_metrics.AverageReturnMetric(buffer_size=10), py_metrics.AverageEpisodeLengthMetric(buffer_size=10), ] # The _train_phase_metrics average over an entire train iteration, # rather than the rolling average of the last 10 episodes. self._train_phase_metrics = [ py_metrics.AverageReturnMetric(name='PhaseAverageReturn', buffer_size=np.inf), py_metrics.AverageEpisodeLengthMetric( name='PhaseAverageEpisodeLength', buffer_size=np.inf), ] self._iteration_metric = py_metrics.CounterMetric( name='Iteration') # Summaries written from python should run every time they are # generated. with tf.compat.v2.summary.record_if(True): self._steps_per_second_ph = tf.compat.v1.placeholder( tf.float32, shape=(), name='steps_per_sec_ph') self._steps_per_second_summary = tf.compat.v2.summary.scalar( name='global_steps_per_sec', data=self._steps_per_second_ph, step=self._global_step) for metric in self._train_metrics: metric.tf_summaries(train_step=self._global_step, step_metrics=self._step_metrics) for metric in self._train_phase_metrics: metric.tf_summaries( train_step=self._global_step, step_metrics=(self._iteration_metric, )) self._iteration_metric.tf_summaries( train_step=self._global_step) if self._do_eval: with self._eval_summary_writer.as_default(): for metric in self._eval_metrics: metric.tf_summaries( train_step=self._global_step, step_metrics=(self._iteration_metric, )) self._train_checkpointer = common.Checkpointer( ckpt_dir=train_dir, agent=agent, global_step=self._global_step, optimizer=optimizer, metrics=metric_utils.MetricsGroup( self._train_metrics + self._train_phase_metrics + [self._iteration_metric], 'train_metrics')) self._policy_checkpointer = common.Checkpointer( ckpt_dir=os.path.join(train_dir, 'policy'), policy=agent.policy, global_step=self._global_step) self._rb_checkpointer = common.Checkpointer( ckpt_dir=os.path.join(train_dir, 'replay_buffer'), max_to_keep=1, replay_buffer=self._replay_buffer) self._init_agent_op = agent.initialize()
def decay(): nonlocal epsilon nonlocal decay_step _epsilon = epsilon epsilon = max(epsilon - epsilon_change, epsilon_min) decay_step += 1 if decay_step % 500 == 0: print('Decaying epsilon from {0} to {1}'.format(_epsilon, epsilon)) return _epsilon return decay eval_policy = agent.policy # Greedy policy # collect_policy = agent.collect_policy # Epsilon-greedy policy collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( agent.collect_policy.wrapped_policy, epsilon=decaying_epsilon()) random_policy = random_tf_policy.RandomTFPolicy( action_spec=collect_policy.action_spec, time_step_spec=collect_policy.time_step_spec) # Random policy # Replay buffer collection replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer( data_spec=agent.collect_data_spec, batch_size=train_env.batch_size, max_length=replay_buffer_max_length) replay_observer = [replay_buffer.add_batch] collect_op = dynamic_step_driver.DynamicStepDriver( train_env, random_policy,
def __init__( self, time_step_spec, action_spec, q_network, optimizer, epsilon_greedy=0.1, # Params for target network updates target_update_tau=1.0, target_update_period=1, # Params for training. td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, gradient_clipping=None, # Params for debugging debug_summaries=False, summarize_grads_and_vars=False): """Creates a DQN Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. q_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type). optimizer: The optimizer to use for training. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of element_wise_huber_loss is used. This function takes as input the target and the estimated Q values and returns the loss for each element of the batch. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. Raises: ValueError: If the action spec contains more than one action. """ flat_action_spec = nest.flatten(action_spec) self._num_actions = [ spec.maximum - spec.minimum + 1 for spec in flat_action_spec ] # TODO(oars): Get DQN working with more than one dim in the actions. if len(flat_action_spec) > 1 or flat_action_spec[0].shape.ndims > 1: raise ValueError('Only one dimensional actions are supported now.') self._q_network = q_network self._target_q_network = self._q_network.copy(name='TargetQNetwork') self._epsilon_greedy = epsilon_greedy self._target_update_tau = target_update_tau self._target_update_period = target_update_period self._optimizer = optimizer self._td_errors_loss_fn = td_errors_loss_fn or element_wise_huber_loss self._gamma = gamma self._reward_scale_factor = reward_scale_factor self._gradient_clipping = gradient_clipping self._target_update_train_op = None policy = q_policy.QPolicy(time_step_spec, action_spec, q_network=self._q_network) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) super(DqnAgent, self).__init__( time_step_spec, action_spec, policy, collect_policy, train_sequence_length=2 if not q_network.state_spec else None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars)
def __init__( self, time_step_spec, action_spec, categorical_q_network, optimizer, min_q_value=-10.0, max_q_value=10.0, epsilon_greedy=0.1, n_step_update=1, boltzmann_temperature=None, # Params for target network updates target_categorical_q_network=None, target_update_tau=1.0, target_update_period=1, # Params for training. td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, gradient_clipping=None, # Params for debugging debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a Categorical DQN Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A `BoundedTensorSpec` representing the actions. categorical_q_network: A categorical_q_network.CategoricalQNetwork that returns the q_distribution for each action. optimizer: The optimizer to use for training. min_q_value: A float specifying the minimum Q-value, used for setting up the support. max_q_value: A float specifying the maximum Q-value, used for setting up the support. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). n_step_update: The number of steps to consider when computing TD error and TD loss. Defaults to single-step updates. Note that this requires the user to call train on Trajectory objects with a time dimension of `n_step_update + 1`. However, note that we do not yet support `n_step_update > 1` in the case of RNNs (i.e., non-empty `q_network.state_spec`). boltzmann_temperature: Temperature value to use for Boltzmann sampling of the actions during data collection. The closer to 0.0, the higher the probability of choosing the best action. target_categorical_q_network: (Optional.) A `tf_agents.network.Network` to be used as the target network during Q learning. Every `target_update_period` train steps, the weights from `categorical_q_network` are copied (possibly with smoothing via `target_update_tau`) to `target_categorical_q_network`. If `target_categorical_q_network` is not provided, it is created by making a copy of `categorical_q_network`, which initializes a new network with the same structure and its own layers and weights. Network copying is performed via the `Network.copy` superclass method, and may inadvertently lead to the resulting network to share weights with the original. This can happen if, for example, the original network accepted a pre-built Keras layer in its `__init__`, or accepted a Keras layer that wasn't built, but neglected to create a new copy. In these cases, it is up to you to provide a target Network having weights that are not shared with the original `categorical_q_network`. If you provide a `target_categorical_q_network` that shares any weights with `categorical_q_network`, a warning will be logged but no exception is thrown. Note; shallow copies of Keras layers may be built via the code: ```python new_layer = type(layer).from_config(layer.get_config()) ``` target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of huber_loss is used. This function takes as input the target and the estimated Q values and returns the loss for each element of the batch. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: TypeError: If the action spec contains more than one action. """ super(CategoricalDqnAgent, self).__init__(time_step_spec, action_spec, categorical_q_network, optimizer, epsilon_greedy=epsilon_greedy, n_step_update=n_step_update, boltzmann_temperature=boltzmann_temperature, target_q_network=target_categorical_q_network, target_update_tau=target_update_tau, target_update_period=target_update_period, td_errors_loss_fn=td_errors_loss_fn, gamma=gamma, reward_scale_factor=reward_scale_factor, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter, name=name) def check_atoms(net, label): num_atoms = getattr(net, 'num_atoms', None) if num_atoms is None: raise TypeError( 'Expected {} to have property `num_atoms`, but it ' 'doesn\'t (note: you likely want to use a ' 'CategoricalQNetwork). Network is: {}'.format(label, net)) return num_atoms num_atoms = check_atoms(self._q_network, 'categorical_q_network') target_num_atoms = check_atoms(self._target_q_network, 'target_categorical_q_network') if num_atoms != target_num_atoms: raise ValueError( 'categorical_q_network and target_categorical_q_network have ' 'different numbers of atoms: {} vs. {}'.format( num_atoms, target_num_atoms)) self._num_atoms = num_atoms self._min_q_value = min_q_value self._max_q_value = max_q_value self._support = tf.linspace(min_q_value, max_q_value, num_atoms) policy = categorical_q_policy.CategoricalQPolicy( min_q_value, max_q_value, self._q_network, self._action_spec) if boltzmann_temperature is not None: self._collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: self._collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) self._policy = greedy_policy.GreedyPolicy(policy) target_policy = categorical_q_policy.CategoricalQPolicy( min_q_value, max_q_value, self._target_q_network, self._action_spec) self._target_greedy_policy = greedy_policy.GreedyPolicy(target_policy)
def __init__(self, time_step_spec, action_spec, actor_network, q_network, actor_optimizer, critic_optimizer, exploration_noise_std=0.1, boltzmann_temperature=None, epsilon_greedy=0.1, q_network_2=None, target_actor_network=None, target_q_network=None, target_q_network_2=None, target_update_tau=1.0, target_update_period=1, actor_update_period=1, dqda_clipping=None, td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, target_policy_noise=0.2, target_policy_noise_clip=0.5, gradient_clipping=None, debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, action_params_mask=None, n_step_update=1, name=None): """Creates a Td3Agent Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A namedtuple of nested BoundedTensorSpec representing the actions. actor_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type). q_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, action, step_type). actor_optimizer: The default optimizer to use for the actor network. critic_optimizer: The default optimizer to use for the critic network. exploration_noise_std: Scale factor on exploration policy noise. q_network_2: (Optional.) A `tf_agents.network.Network` to be used as the second critic network during Q learning. The weights from `q_network` are copied if this is not provided. target_actor_network: (Optional.) A `tf_agents.network.Network` to be used as the target actor network during Q learning. Every `target_update_period` train steps, the weights from `actor_network` are copied (possibly withsmoothing via `target_update_tau`) to ` target_actor_network`. If `target_actor_network` is not provided, it is created by making a copy of `actor_network`, which initializes a new network with the same structure and its own layers and weights. Performing a `Network.copy` does not work when the network instance already has trainable parameters (e.g., has already been built, or when the network is sharing layers with another). In these cases, it is up to you to build a copy having weights that are not shared with the original `actor_network`, so that this can be used as a target network. If you provide a `target_actor_network` that shares any weights with `actor_network`, a warning will be logged but no exception is thrown. target_q_network: (Optional.) Similar network as target_actor_network but for the q_network. See documentation for target_actor_network. target_q_network_2: (Optional.) Similar network as target_actor_network but for the q_network_2. See documentation for target_actor_network. Will only be used if 'q_network_2' is also specified. target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. actor_update_period: Period for the optimization step on actor network. dqda_clipping: A scalar or float clips the gradient dqda element-wise between [-dqda_clipping, dqda_clipping]. Default is None representing no clippiing. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of elementwise huber_loss is used. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. target_policy_noise: Scale factor on target action noise target_policy_noise_clip: Value to clip noise. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. action_params_mask: A mask of continuous parameter actions for discrete action name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. """ tf.Module.__init__(self, name=name) self._actor_network = actor_network self._target_actor_network = common.maybe_copy_target_network_with_checks( self._actor_network, target_actor_network, 'TargetActorNetwork') # critic network here is Q-network self._q_network_1 = q_network self._target_q_network_1 = ( common.maybe_copy_target_network_with_checks( self._q_network_1, target_q_network, 'TargetCriticNetwork1')) if q_network_2 is not None: self._q_network_2 = q_network_2 else: self._q_network_2 = q_network.copy(name='CriticNetwork2') # Do not use target_q_network_2 if q_network_2 is None. target_q_network_2 = None self._target_q_network_2 = ( common.maybe_copy_target_network_with_checks( self._q_network_2, target_q_network_2, 'TargetCriticNetwork2')) self._actor_optimizer = actor_optimizer self._critic_optimizer = critic_optimizer self._exploration_noise_std = exploration_noise_std self._epsilon_greedy = epsilon_greedy self._boltzmann_temperature = boltzmann_temperature self._target_update_tau = target_update_tau self._target_update_period = target_update_period self._actor_update_period = actor_update_period self._dqda_clipping = dqda_clipping self._td_errors_loss_fn = (td_errors_loss_fn or common.element_wise_huber_loss) self._gamma = gamma self._reward_scale_factor = reward_scale_factor self._target_policy_noise = target_policy_noise self._target_policy_noise_clip = target_policy_noise_clip self._gradient_clipping = gradient_clipping self._update_target = self._get_target_updater(target_update_tau, target_update_period) policy = actor_policy.ActorPolicy( time_step_spec=time_step_spec, action_spec=action_spec.actor_network, actor_network=self._actor_network, clip=True) policy = mixed_q_policy.MixedQPolicy(policy, time_step_spec=time_step_spec, action_spec=action_spec.q_network, q_network=q_network) collect_policy = actor_policy.ActorPolicy( time_step_spec=time_step_spec, action_spec=action_spec.actor_network, actor_network=self._actor_network, clip=False) collect_policy = gaussian_policy.GaussianPolicy( collect_policy, scale=self._exploration_noise_std, clip=True) collect_policy = mixed_q_policy.MixedQPolicy( collect_policy, time_step_spec=time_step_spec, action_spec=action_spec.q_network, q_network=q_network) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( collect_policy, temperature=self._boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( collect_policy, epsilon=self._epsilon_greedy) # Create self._target_greedy_policy in order to compute target Q-values. target_policy = actor_policy.ActorPolicy( time_step_spec=time_step_spec, action_spec=action_spec.actor_network, actor_network=self._target_actor_network, clip=True) target_policy = mixed_q_policy.MixedQPolicy( target_policy, time_step_spec=time_step_spec, action_spec=action_spec.q_network, q_network=self._target_q_network_1) self._target_greedy_policy = greedy_policy.GreedyPolicy(target_policy) self._action_params_mask = action_params_mask self._n_step_update = n_step_update if action_spec.actor_network is not None and action_params_mask is None: raise ValueError( "action_params_mask is required for actor network") super(MixedTd3Agent, self).__init__(time_step_spec, action_spec, policy, collect_policy, train_sequence_length=2 if not self._actor_network.state_spec else None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter)
def __init__( self, time_step_spec, action_spec, reward_network, optimizer, epsilon, observation_and_action_constraint_splitter=None, # Params for training. error_loss_fn=tf.compat.v1.losses.mean_squared_error, gradient_clipping=None, # Params for debugging. debug_summaries=False, summarize_grads_and_vars=False, enable_summaries=True, expose_predicted_rewards=False, train_step_counter=None, name=None): """Creates a Neural Epsilon Greedy Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. reward_network: A `tf_agents.network.Network` to be used by the agent. The network will be called with call(observation, step_type) and it is expected to provide a reward prediction for all actions. *Note*: when using `observation_and_action_constraint_splitter`, make sure the `reward_network` is compatible with the network-specific half of the output of the `observation_and_action_constraint_splitter`. In particular, `observation_and_action_constraint_splitter` will be called on the observation before passing to the network. optimizer: The optimizer to use for training. epsilon: A float representing the probability of choosing a random action instead of the greedy action. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. error_loss_fn: A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is `tf.losses.mean_squared_error`. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. enable_summaries: A Python bool, default True. When False, all summaries (debug or otherwise) should not be written. expose_predicted_rewards: (bool) Whether to expose the predicted rewards in the policy info field under the name 'predicted_rewards'. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0. """ super(NeuralEpsilonGreedyAgent, self).__init__(time_step_spec=time_step_spec, action_spec=action_spec, reward_network=reward_network, optimizer=optimizer, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), error_loss_fn=error_loss_fn, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, enable_summaries=enable_summaries, expose_predicted_rewards=expose_predicted_rewards, train_step_counter=train_step_counter, name=name) self._policy = epsilon_greedy_policy.EpsilonGreedyPolicy( self._policy, epsilon=epsilon) self._collect_policy = self._policy
def __init__( self, time_step_spec, action_spec, cloning_network, optimizer, epsilon_greedy=0.1, # Params for training. loss_fn=None, gradient_clipping=None, # Params for debugging debug_summaries=False, summarize_grads_and_vars=False): """Creates an behavioral cloning Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. cloning_network: A tf_agents.network.Network to be used by the agent. The network will be called as ``` network(observation, step_type, network_state=None) ``` (with `network_state` optional) and must return a 2-tuple with elements `(output, next_network_state)` where `output` will be passed as the first argument to `loss_fn`, and used by a `Policy`. Input tensors will be shaped `[batch, time, ...]` when training, and they will be shaped `[batch, ...]` when the network is called within a `Policy`. If `cloning_network` has an empty network state, then for training `time` will always be `1` (individual examples). optimizer: The optimizer to use for training. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). loss_fn: A function for computing the error between the output of the cloning network and the action that was taken. If None, the loss depends on the action dtype. If the dtype is integer, then `loss_fn` is ```python def loss_fn(logits, action): return tf.nn.sparse_softmax_cross_entropy_with_logits( labels=action - action_spec.minimum, logits=logits) ``` If the dtype is floating point, the loss is `tf.math.squared_difference`. `loss_fn` must return a loss value for each element of the batch. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. Raises: NotImplementedError: If the action spec contains more than one action. """ flat_action_spec = nest.flatten(action_spec) self._num_actions = [ spec.maximum - spec.minimum + 1 for spec in flat_action_spec ] # TODO(oars): Get behavioral cloning working with more than one dim in # the actions. if len(flat_action_spec) > 1: raise NotImplementedError( 'Multi-arity actions are not currently supported.') if flat_action_spec[0].dtype.is_floating: if loss_fn is None: loss_fn = tf.math.squared_difference else: if flat_action_spec[0].shape.ndims > 1: raise NotImplementedError( 'Only scalar and one dimensional integer actions are supported.' ) if loss_fn is None: # TODO(ebrevdo): Maybe move the subtraction of the minimum into a # self._label_fn and rewrite this. def xent_loss_fn(logits, actions): # Subtract the minimum so that we get a proper cross entropy loss on # [0, maximum - minimum). return tf.nn.sparse_softmax_cross_entropy_with_logits( logits=logits, labels=actions - flat_action_spec[0].minimum) loss_fn = xent_loss_fn self._cloning_network = cloning_network self._loss_fn = loss_fn self._epsilon_greedy = epsilon_greedy self._optimizer = optimizer self._gradient_clipping = gradient_clipping policy = q_policy.QPolicy(time_step_spec, action_spec, q_network=self._cloning_network) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) super(BehavioralCloningAgent, self).__init__(time_step_spec, action_spec, policy, collect_policy, train_sequence_length=1 if not cloning_network.state_spec else None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars)
def _setup_policy(self, time_step_spec, action_spec): policy = Policy(time_step_spec, action_spec, network=self._network) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) return policy, collect_policy
def __init__( self, time_step_spec, action_spec, q_network, optimizer, epsilon_greedy=0.1, boltzmann_temperature=None, # Params for target network updates target_update_tau=1.0, target_update_period=1, # Params for training. td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, gradient_clipping=None, # Params for debugging debug_summaries=False, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a DQN Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. q_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type). optimizer: The optimizer to use for training. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). boltzmann_temperature: Temperature value to use for Boltzmann sampling of the actions during data collection. The closer to 0.0, the higher the probability of choosing the best action. target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of element_wise_huber_loss is used. This function takes as input the target and the estimated Q values and returns the loss for each element of the batch. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or action spec minimum is not equal to 0. """ tf.Module.__init__(self, name=name) flat_action_spec = tf.nest.flatten(action_spec) self._num_actions = [ spec.maximum - spec.minimum + 1 for spec in flat_action_spec ] # TODO(oars): Get DQN working with more than one dim in the actions. if len(flat_action_spec) > 1 or flat_action_spec[0].shape.ndims > 1: raise ValueError('Only one dimensional actions are supported now.') if not all(spec.minimum == 0 for spec in flat_action_spec): raise ValueError( 'Action specs should have minimum of 0, but saw: {0}'.format( [spec.minimum for spec in flat_action_spec])) if epsilon_greedy is not None and boltzmann_temperature is not None: raise ValueError( 'Configured both epsilon_greedy value {} and temperature {}, ' 'however only one of them can be used for exploration.'.format( epsilon_greedy, boltzmann_temperature)) self._q_network = q_network self._target_q_network = self._q_network.copy(name='TargetQNetwork') self._epsilon_greedy = epsilon_greedy self._boltzmann_temperature = boltzmann_temperature self._optimizer = optimizer self._td_errors_loss_fn = td_errors_loss_fn or element_wise_huber_loss self._gamma = gamma self._reward_scale_factor = reward_scale_factor self._gradient_clipping = gradient_clipping self._update_target = self._get_target_updater( target_update_tau, target_update_period) policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._q_network) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) super(DqnAgent, self).__init__( time_step_spec, action_spec, policy, collect_policy, train_sequence_length=2 if not q_network.state_spec else None, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter)
def __init__( self, time_step_spec, action_spec, q_network, optimizer, epsilon_greedy=0.1, n_step_update=1, boltzmann_temperature=None, emit_log_probability=False, update_period=None, # Params for target network updates target_update_tau=1.0, target_update_period=1, # Params for training. td_errors_loss_fn=None, gamma=1.0, reward_scale_factor=1.0, gradient_clipping=None, # Params for debugging debug_summaries=False, enable_functions=True, summarize_grads_and_vars=False, train_step_counter=None, name=None): """Creates a DQN Agent. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of BoundedTensorSpec representing the actions. q_network: A tf_agents.network.Network to be used by the agent. The network will be called with call(observation, step_type). optimizer: The optimizer to use for training. epsilon_greedy: probability of choosing a random action in the default epsilon-greedy collect policy (used only if a wrapper is not provided to the collect_policy method). n_step_update: The number of steps to consider when computing TD error and TD loss. Defaults to single-step updates. Note that this requires the user to call train on Trajectory objects with a time dimension of `n_step_update + 1`. However, note that we do not yet support `n_step_update > 1` in the case of RNNs (i.e., non-empty `q_network.state_spec`). boltzmann_temperature: Temperature value to use for Boltzmann sampling of the actions during data collection. The closer to 0.0, the higher the probability of choosing the best action. emit_log_probability: Whether policies emit log probabilities or not. update_period: Update period. target_update_tau: Factor for soft update of the target networks. target_update_period: Period for soft update of the target networks. td_errors_loss_fn: A function for computing the TD errors loss. If None, a default value of element_wise_huber_loss is used. This function takes as input the target and the estimated Q values and returns the loss for each element of the batch. gamma: A discount factor for future rewards. reward_scale_factor: Multiplicative scale for the reward. gradient_clipping: Norm length to clip gradients. debug_summaries: A bool to gather debug summaries. enable_functions: A bool to decide whether or not to enable tf function summarize_grads_and_vars: If True, gradient and network variable summaries will be written during training. train_step_counter: An optional counter to increment every time the train op is run. Defaults to the global_step. name: The name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or action spec minimum is not equal to 0. NotImplementedError: If `q_network` has non-empty `state_spec` (i.e., an RNN is provided) and `n_step_update > 1`. """ tf.Module.__init__(self, name=name) flat_action_spec = tf.nest.flatten(action_spec) self._num_actions = [ spec.maximum - spec.minimum + 1 for spec in flat_action_spec ] if len(flat_action_spec) > 1 or flat_action_spec[0].shape.ndims > 1: raise ValueError('Only one dimensional actions are supported now.') if not all(spec.minimum == 0 for spec in flat_action_spec): raise ValueError( 'Action specs should have minimum of 0, but saw: {0}'.format( [spec.minimum for spec in flat_action_spec])) if epsilon_greedy is not None and boltzmann_temperature is not None: raise ValueError( 'Configured both epsilon_greedy value {} and temperature {}, ' 'however only one of them can be used for exploration.'.format( epsilon_greedy, boltzmann_temperature)) self._q_network = q_network self._target_q_network = self._q_network.copy(name='TargetQNetwork') self._epsilon_greedy = epsilon_greedy self._n_step_update = n_step_update self._boltzmann_temperature = boltzmann_temperature self._optimizer = optimizer self._td_errors_loss_fn = td_errors_loss_fn or element_wise_huber_loss self._gamma = gamma self._reward_scale_factor = reward_scale_factor self._gradient_clipping = gradient_clipping self._update_target = self._get_target_updater(target_update_tau, target_update_period) policy = q_policy.QPolicy( time_step_spec, action_spec, q_network=self._q_network, emit_log_probability=emit_log_probability) if boltzmann_temperature is not None: collect_policy = boltzmann_policy.BoltzmannPolicy( policy, temperature=self._boltzmann_temperature) else: collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( policy, epsilon=self._epsilon_greedy) policy = greedy_policy.GreedyPolicy(policy) if q_network.state_spec and n_step_update != 1: raise NotImplementedError( 'DqnAgent does not currently support n-step updates with stateful ' 'networks (i.e., RNNs), but n_step_update = {}'.format(n_step_update)) train_sequence_length = ( n_step_update + 1 if not q_network.state_spec else None) super(DqnAgent, self).__init__( time_step_spec, action_spec, policy, collect_policy, train_sequence_length=train_sequence_length, update_period=update_period, debug_summaries=debug_summaries, enable_functions=enable_functions, summarize_grads_and_vars=summarize_grads_and_vars, train_step_counter=train_step_counter) tf.compat.v1.summary.scalar( 'epsilon/' + self.name, self._epsilon_greedy, collections=['train_' + self.name])
def __init__( self, time_step_spec: types.TimeStep, action_spec: types.BoundedTensorSpec, reward_network: types.Network, optimizer: types.Optimizer, epsilon: float, observation_and_action_constraint_splitter: Optional[ types.Splitter] = None, accepts_per_arm_features: bool = False, constraints: Iterable[constr.NeuralConstraint] = (), # Params for training. error_loss_fn: types.LossFn = tf.compat.v1.losses. mean_squared_error, gradient_clipping: Optional[float] = None, # Params for debugging. debug_summaries: bool = False, summarize_grads_and_vars: bool = False, enable_summaries: bool = True, emit_policy_info: Tuple[Text, ...] = (), train_step_counter: Optional[tf.Variable] = None, laplacian_matrix: Optional[types.Float] = None, laplacian_smoothing_weight: float = 0.001, info_fields_to_inherit_from_greedy: Sequence[Text] = (), name: Optional[Text] = None): """Creates a Neural Epsilon Greedy Agent. For more details about the Laplacian smoothing regularization, please see the documentation of the `GreedyRewardPredictionAgent`. Args: time_step_spec: A `TimeStep` spec of the expected time_steps. action_spec: A nest of `BoundedTensorSpec` representing the actions. reward_network: A `tf_agents.network.Network` to be used by the agent. The network will be called with call(observation, step_type) and it is expected to provide a reward prediction for all actions. *Note*: when using `observation_and_action_constraint_splitter`, make sure the `reward_network` is compatible with the network-specific half of the output of the `observation_and_action_constraint_splitter`. In particular, `observation_and_action_constraint_splitter` will be called on the observation before passing to the network. optimizer: The optimizer to use for training. epsilon: A float representing the probability of choosing a random action instead of the greedy action. observation_and_action_constraint_splitter: A function used for masking valid/invalid actions with each state of the environment. The function takes in a full observation and returns a tuple consisting of 1) the part of the observation intended as input to the bandit agent and policy, and 2) the boolean mask. This function should also work with a `TensorSpec` as input, and should output `TensorSpec` objects for the observation and mask. accepts_per_arm_features: (bool) Whether the policy accepts per-arm features. constraints: iterable of constraints objects that are instances of `tf_agents.bandits.agents.NeuralConstraint`. error_loss_fn: A function for computing the error loss, taking parameters labels, predictions, and weights (any function from tf.losses would work). The default is `tf.losses.mean_squared_error`. gradient_clipping: A float representing the norm length to clip gradients (or None for no clipping.) debug_summaries: A Python bool, default False. When True, debug summaries are gathered. summarize_grads_and_vars: A Python bool, default False. When True, gradients and network variable summaries are written during training. enable_summaries: A Python bool, default True. When False, all summaries (debug or otherwise) should not be written. emit_policy_info: (tuple of strings) what side information we want to get as part of the policy info. Allowed values can be found in `policy_utilities.PolicyInfo`. train_step_counter: An optional `tf.Variable` to increment every time the train op is run. Defaults to the `global_step`. laplacian_matrix: A float `Tensor` shaped `[num_actions, num_actions]`. This holds the Laplacian matrix used to regularize the smoothness of the estimated expected reward function. This only applies to problems where the actions have a graph structure. If `None`, the regularization is not applied. laplacian_smoothing_weight: A float that determines the weight of the regularization term. Note that this has no effect if `laplacian_matrix` above is `None`. info_fields_to_inherit_from_greedy: List of info fields that are reported from the greedy policy even when exploratory action is taken. name: Python str name of this agent. All variables in this module will fall under that name. Defaults to the class name. Raises: ValueError: If the action spec contains more than one action or or it is not a bounded scalar int32 spec with minimum 0. """ super(NeuralEpsilonGreedyAgent, self).__init__( time_step_spec=time_step_spec, action_spec=action_spec, reward_network=reward_network, optimizer=optimizer, observation_and_action_constraint_splitter=( observation_and_action_constraint_splitter), accepts_per_arm_features=accepts_per_arm_features, constraints=constraints, error_loss_fn=error_loss_fn, gradient_clipping=gradient_clipping, debug_summaries=debug_summaries, summarize_grads_and_vars=summarize_grads_and_vars, enable_summaries=enable_summaries, emit_policy_info=emit_policy_info, train_step_counter=train_step_counter, laplacian_matrix=laplacian_matrix, laplacian_smoothing_weight=laplacian_smoothing_weight, name=name) self._policy = epsilon_greedy_policy.EpsilonGreedyPolicy( self._policy, epsilon=epsilon, info_fields_to_inherit_from_greedy= info_fields_to_inherit_from_greedy) self._collect_policy = self._policy
def train(): global VERBOSE environment = TradeEnvironment() # utils.validate_py_environment(environment, episodes=5) # Environments train_env = tf_py_environment.TFPyEnvironment(environment) eval_env = tf_py_environment.TFPyEnvironment(environment) num_iterations = 50 fc_layer_params = (512, ) # ~ (17 + 1001) / 2 input_fc_layer_params = (17, ) output_fc_layer_params = (20, ) lstm_size = (17, ) initial_collect_steps = 20 collect_steps_per_iteration = 1 batch_size = 64 replay_buffer_capacity = 10000 gamma = 0.99 # check if 1 will work here target_update_tau = 0.05 target_update_period = 5 epsilon_greedy = 0.1 reward_scale_factor = 1.0 learning_rate = 1e-2 log_interval = 30 num_eval_episodes = 5 eval_interval = 15 # q_net = q_network.QNetwork( # train_env.observation_spec(), # train_env.action_spec(), # fc_layer_params=fc_layer_params, # ) q_net = q_rnn_network.QRnnNetwork( train_env.observation_spec(), train_env.action_spec(), input_fc_layer_params=input_fc_layer_params, lstm_size=lstm_size, output_fc_layer_params=output_fc_layer_params, ) optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate) train_step_counter = tf.compat.v2.Variable(0) tf_agent = dqn_agent.DqnAgent( train_env.time_step_spec(), train_env.action_spec(), q_network=q_net, optimizer=optimizer, epsilon_greedy=epsilon_greedy, target_update_tau=target_update_tau, target_update_period=target_update_period, gamma=gamma, reward_scale_factor=reward_scale_factor, td_errors_loss_fn=dqn_agent.element_wise_squared_loss, train_step_counter=train_step_counter, gradient_clipping=None, debug_summaries=False, summarize_grads_and_vars=False, ) q_policy = FilteredQPolicy( tf_agent._time_step_spec, tf_agent._action_spec, q_network=tf_agent._q_network, ) # Valid policy to pre-fill replay buffer dummy_policy = DummyTradePolicy( train_env.time_step_spec(), train_env.action_spec(), ) # Main agent's policy; greedy one policy = greedy_policy.GreedyPolicy(q_policy) filtered_random_py_policy = FilteredRandomPyPolicy( time_step_spec=policy.time_step_spec, action_spec=policy.action_spec, ) filtered_random_tf_policy = tf_py_policy.TFPyPolicy( filtered_random_py_policy) collect_policy = epsilon_greedy_policy.EpsilonGreedyPolicy( q_policy, epsilon=tf_agent._epsilon_greedy) # Patch random policy for epsilon greedy collect policy filtered_random_tf_policy = FilteredRandomTFPolicy( time_step_spec=policy.time_step_spec, action_spec=policy.action_spec, ) collect_policy._random_policy = filtered_random_tf_policy tf_agent._policy = policy tf_agent._collect_policy = collect_policy tf_agent.initialize() replay_buffer = tf_uniform_replay_buffer.TFUniformReplayBuffer( data_spec=tf_agent.collect_data_spec, batch_size=train_env.batch_size, max_length=replay_buffer_capacity, ) print( 'Pre-filling replay buffer in {} steps'.format(initial_collect_steps)) for _ in range(initial_collect_steps): traj = collect_step(train_env, dummy_policy) replay_buffer.add_batch(traj) dataset = replay_buffer.as_dataset( num_parallel_calls=3, sample_batch_size=batch_size, num_steps=2, ).prefetch(3) iterator = iter(dataset) # Train tf_agent.train = common.function(tf_agent.train) tf_agent.train_step_counter.assign(0) avg_return = compute_avg_return(eval_env, tf_agent.policy, num_eval_episodes) returns = [avg_return] print('Starting iterations...') for i in range(num_iterations): # fill replay buffer for _ in range(collect_steps_per_iteration): traj = collect_step(train_env, tf_agent.collect_policy) # Add trajectory to the replay buffer replay_buffer.add_batch(traj) experience, _ = next(iterator) train_loss = tf_agent.train(experience) step = tf_agent.train_step_counter.numpy() if step % log_interval == 0: print('step = {0}: loss = {1}'.format(step, train_loss.loss)) if step % eval_interval == 0: avg_return = compute_avg_return(eval_env, tf_agent.policy, num_eval_episodes) print('step = {0}: avg return = {1}'.format(step, avg_return)) returns.append(avg_return) print('Finished {} iterations!'.format(num_iterations)) print('Playing with resulting policy') VERBOSE = True r = compute_avg_return(eval_env, tf_agent.policy, 1) print('Result: {}'.format(r)) steps = range(0, num_iterations + 1, eval_interval) # merged = tf.summary.merge_all() # writer = tf.summary.FileWriter(FLAGS.log_dir) # # writer.close() print('Check out chart for learning') plt.plot(steps, returns) plt.ylabel('Average Return') plt.xlabel('Step') plt.ylim(top=1000) plt.show()